uncertainty quantification of climate sensitivity: state ...€¦ · uncertainty quantification of...

Department of Physics

Uncertainty quantification of climate sensitivity: State-dependence, extreme values and the probability of tipping

Anna von der Heydt1,2

1. Institute for Marine and Atmospheric Research, Department of Physics, Utrecht University, Utrecht, The Netherlands.

2. Centre for Complex Systems Studies, Utrecht University, Utrecht, The Netherlands.

Dr. Anna von der Heydt

Climate forcing…

2


… Models’ Response

3

Shaded area includes results of > 20 climate models (GCMs)


Tipping elements in the climate system

4

Need a comprehensive framework to quantify the Climate Response!

to the CO2 fertilization effect or decreasing uptake due to a de-crease in rainfall). For some of the tipping elements, crossing thetipping point could trigger an abrupt, nonlinear response (e.g.,conversion of large areas of the Amazon rainforest to a savanna orseasonally dry forest), while for others, crossing the tipping pointwould lead to a more gradual but self-perpetuating response(large-scale loss of permafrost). There could also be considerablelags after the crossing of a threshold, particularly for those tippingelements that involve the melting of large masses of ice. However,in some cases, ice loss can be very rapid when occurring asmassive iceberg outbreaks (e.g., Heinrich Events).

For some feedback processes, the magnitude—and even thedirection—depend on the rate of climate change. If the rate ofclimate change is small, the shift in biomes can track the change intemperature/moisture, and the biomes may shift gradually, po-tentially taking up carbon from the atmosphere as the climate warmsand atmospheric CO2 concentration increases. However, if the rate ofclimate change is too large or too fast, a tipping point can be crossed,and a rapid biome shift may occur via extensive disturbances (e.g.,wildfires, insect attacks, droughts) that can abruptly remove anexisting biome. In some terrestrial cases, such as widespread wild-fires, there could be a pulse of carbon to the atmosphere, which iflarge enough, could influence the trajectory of the Earth System (29).

Varying response rates to a changing climate could lead tocomplex biosphere dynamics with implications for feedbackprocesses. For example, delays in permafrost thawing would mostlikely delay the projected northward migration of boreal forests(30), while warming of the southern areas of these forests couldresult in their conversion to steppe grasslands of significantlylower carbon storage capacity. The overall result would be apositive feedback to the climate system.

The so-called “greening” of the planet, caused by enhancedplant growth due to increasing atmospheric CO2 concentration(31), has increased the land carbon sink in recent decades (32).However, increasing atmospheric CO2 raises temperature, andhotter leaves photosynthesize less well. Other feedbacks are alsoinvolved—for instance, warming the soil increases microbial res-piration, releasing CO2 back into the atmosphere.

Our analysis focuses on the strength of the feedback betweennow and 2100. However, several of the feedbacks that shownegligible or very small magnitude by 2100 could nevertheless betriggered well before then, and they could eventually generatesignificant feedback strength over longer timeframes—centuriesand even millennia—and thus, influence the long-term trajectoryof the Earth System. These feedback processes include perma-frost thawing, decomposition of ocean methane hydrates, in-creased marine bacterial respiration, and loss of polar ice sheetsaccompanied by a rise in sea levels and potential amplification oftemperature rise through changes in ocean circulation (33).

Tipping Cascades. Fig. 3 shows a global map of some potentialtipping cascades. The tipping elements fall into three clustersbased on their estimated threshold temperature (12, 17, 39).Cascades could be formed when a rise in global temperaturereaches the level of the lower-temperature cluster, activatingtipping elements, such as loss of the Greenland Ice Sheet or Arcticsea ice. These tipping elements, along with some of the non-tipping element feedbacks (e.g., gradual weakening of land andocean physiological carbon sinks), could push the global averagetemperature even higher, inducing tipping in mid- and higher-temperature clusters. For example, tipping (loss) of the Green-land Ice Sheet could trigger a critical transition in the AtlanticMeridional Ocean Circulation (AMOC), which could together, bycausing sea-level rise and Southern Ocean heat accumulation,accelerate ice loss from the East Antarctic Ice Sheet (32, 40) ontimescales of centuries (41).

Observations of past behavior support an important contri-bution of changes in ocean circulation to such feedback cascades.During previous glaciations, the climate system flickered betweentwo states that seem to reflect changes in convective activity in theNordic seas and changes in the activity of the AMOC. Thesevariations caused typical temperature response patterns called the“bipolar seesaw” (42–44). During extremely cold conditions in thenorth, heat accumulated in the Southern Ocean, and Antarcticawarmed. Eventually, the heat made its way north and generatedsubsurface warming that may have been instrumental in destabi-lizing the edges of the Northern Hemisphere ice sheets (45).

If Greenland and the West Antarctic Ice Sheet melt in the fu-ture, the freshening and cooling of nearby surface waters will havesignificant effects on the ocean circulation. While the probabilityof significant circulation changes is difficult to quantify, climatemodel simulations suggest that freshwater inputs compatible withcurrent rates of Greenland melting are sufficient to have mea-surable effects on ocean temperature and circulation (46, 47).Sustained warming of the northern high latitudes as a result of thisprocess could accelerate feedbacks or activate tipping elementsin that region, such as permafrost degradation, loss of Arctic seaice, and boreal forest dieback.

While this may seem to be an extreme scenario, it illustratesthat a warming into the range of even the lower-temperaturecluster (i.e., the Paris targets) could lead to tipping in the mid- andhigher-temperature clusters via cascade effects. Based on thisanalysis of tipping cascades and taking a risk-averse approach, wesuggest that a potential planetary threshold could occur at atemperature rise as low as !2.0 °C above preindustrial (Fig. 1).

Alternative Stabilized Earth PathwayIf the world’s societies want to avoid crossing a potential thresholdthat locks the Earth System into the Hothouse Earth pathway, thenit is critical that they make deliberate decisions to avoid this risk

Fig. 3. Global map of potential tipping cascades. The individualtipping elements are color- coded according to estimated thresholdsin global average surface temperature (tipping points) (12, 34).Arrows show the potential interactions among the tipping elementsbased on expert elicitation that could generate cascades. Note that,although the risk for tipping (loss of) the East Antarctic Ice Sheet isproposed at>5 °C, somemarine-based sectors in East Antarctica maybe vulnerable at lower temperatures (35–38).

Steffen et al. PNAS | August 14, 2018 | vol. 115 | no. 33 | 8255

Steffen et al. PNAS 2018


! Equilibrium change in global mean surface temperature after a doubling of the atmospheric CO2 concentration.

What is climate sensitivity?

5

Doubling of pCO2

Radiative perturbation

ΔR0

Climate System (no feedbacks) ΔT = S0 ΔR0

! No feedbacks:

Planck response S0= 0.3 K/(W/m2)


! Equilibrium change in global mean surface temperature after a doubling of the atmospheric CO2 concentration.

! With feedbacks:

What is climate sensitivity?

5

Doubling of pCO2

Radiative perturbation

ΔR0

Climate System (no feedbacks) ΔT = S0 ΔR0

ΔRfb = c1 ΔT

Climate System with feedbacks

ΔT = S0 (ΔR0+c1ΔT) = S ΔR0

S = ΔT/ΔR0

! No feedbacks:

Planck response S0= 0.3 K/(W/m2)


! Timescales and equilibrium Slow and fast feedback processes. Timescale separation.

! Dependence on the background climate (Fast) feedback processes are not “constant”.

! Tipping points in the climate system New ‘flavours’ of climate sensitivity. Extremes in climate sensitivity vs probability of tipping.

Quantifying climate sensitivity: problems

6


Time scales & equilibrium

7

CH4

Years Decades Centuries Millennia Multi-millennia /Myr

Upper ocean

VegetationDust/aerosol

Entire oceans

Clouds, water vapour, Lapse rate, snow/sea ice

Land ice sheets

Carbon cycle

Weathering

Plate tectonics

slow feedbacksfast feedbacks

~ 100 years

PALAEOSENS Rohling et al., Nature 491 (2012) ↖

[15]. Section 4 finishes with a discussion of conclusions and some challenges for the future.

2 Sensitivities and the climate attractorsec:sens

In order to understand variability, abrupt transitions and response to perturbations we con-sider the climate system as a high-dimensional multiscale complex dynamical system whoseevolving trajectories form a climate attractor. The ECS can be defined on this attractor andregimes or states may be identified where a linear approximation of the response may bereasonable. Tipping points visible in the GMST will show up as large but occasional shiftsbetween di↵erent “climate regimes” of the attractor, or indeed di↵erent attractors. We vi-sualise the attractor by projection onto climate variables relevant for determining ECS, i.e.the GMST T and the radiative forcing R [32]. Consider the energy balance equation

cTdT

dt= Rforcing +Rslow +Rfast �ROLW , (1) e:energybalance

where the left hand side represents the rate of change of the global mean surface temperatureT (with a heat capacity cT ) and on the right hand side Rforcing is the (external) radiativeforcing (including changes in CO2), Rslow (Rfast) is the radiative perturbation due to allslow (fast) feedback processes within the climate system and ROLW is the outgoing longwaveradiation, respectively. Following the formalism of [26], the specific climate sensitivity is

Sforcing,slow =�T

�Rforcing +�Rslow⇡ dT

d(Rforcing +Rslow), (2) eq:Sspecific

which equals the Charney sensitivity S if �Rslow is the sum of all slow feedback processescontributing to the ECS (and under the assumption of time scale separation). In practise,only some of the slow processes are accessible from palaeoclimate records (e.g. only landice), in which case the specific climate sensitivity is only an approximation of the Charneysensitivity [26] (e.g. S[CO2,LI] is the specific climate sensitivity considering only land icechanges as slow feedback). The use of the specific climate sensitivity is that it can give alinear prediction

T0 = T + Sforcing,slow (�Rforcing +�Rslow) . (3) eq:linpred

For a specific energy balance model including regime shifts we can explicitly calculate ECSfor the di↵erent regimes, see section 3. We note that several other authors have highlightedthe need to improved notes of ECS: these include [7] who propose to use a measure-basedapproach to understand climate sensitivity and [11] who consider conditional climate sensi-tivities constrained by temperature, coupled with resilience measures for switching to otherregimes.

2.1 Observation of the climate attractor

We consider the climate system as a high dimensional dynamical system that evolves alongtrajectories x(t) according to a smooth flow

x(t) = 't(x(0)) (4) e:climatesystem

4

Earth system sensitivity

‘Correct’ for slow feedbacks, e.g.

‘Equilibrium’ sensitivity S:

This finally leads to the expressions for the specific climate sensitivities

S[CO2] =�T

�R[CO2]

=��T

�R[OLW ] +�R[SI] +�R[surf ] +�R[LI]

(21)

S[CO2,LI] =�T

�R[CO2] +�R[LI]

=��T

�R[OLW ] +�R[SI] +�R[surf ]

(22)

S[CO2,LI,SI] =�T

�R[CO2] +�R[LI] +�R[SI]

=��T

�R[OLW ] +�R[surf ]

. (23)

The last expression should approximate the sensitivity without feedbacks (i.e. only Planck

feedback), S0 = (�4"�BT3)�1 ' 0.3 K (W m

�2)�1. In the model there is, however, one

more radiation term due to the atmosphere-ocean heat exchange (�Rsurf ), which acts on

fast to intermediate time scales. Therefore, S[CO2,LI,SI] still slightly deviates from the Planck

sensitivity.

Sp

= S[CO2] =�T

�R[CO2]

(24)

S[CO2,LI] =�T

�R[CO2] +�R[LI]

(25)

3

[15]. Section 4 finishes with a discussion of conclusions and some challenges for the future.

2 Sensitivities and the climate attractorsec:sens

In order to understand variability, abrupt transitions and response to perturbations we con-sider the climate system as a high-dimensional multiscale complex dynamical system whoseevolving trajectories form a climate attractor. The ECS can be defined on this attractor andregimes or states may be identified where a linear approximation of the response may bereasonable. Tipping points visible in the GMST will show up as large but occasional shiftsbetween di↵erent “climate regimes” of the attractor, or indeed di↵erent attractors. We vi-sualise the attractor by projection onto climate variables relevant for determining ECS, i.e.the GMST T and the radiative forcing R [32]. Consider the energy balance equation

cTdT

dt= Rforcing +Rslow +Rfast �ROLW , (1) e:energybalance

where the left hand side represents the rate of change of the global mean surface temperatureT (with a heat capacity cT ) and on the right hand side Rforcing is the (external) radiativeforcing (including changes in CO2), Rslow (Rfast) is the radiative perturbation due to allslow (fast) feedback processes within the climate system and ROLW is the outgoing longwaveradiation, respectively. Following the formalism of [26], the specific climate sensitivity is

Sforcing,slow =�T

�Rforcing +�Rslow⇡ dT

d(Rforcing +Rslow), (2) eq:Sspecific

which equals the Charney sensitivity S if �Rslow is the sum of all slow feedback processescontributing to the ECS (and under the assumption of time scale separation). In practise,only some of the slow processes are accessible from palaeoclimate records (e.g. only landice), in which case the specific climate sensitivity is only an approximation of the Charneysensitivity [26] (e.g. S[CO2,LI] is the specific climate sensitivity considering only land icechanges as slow feedback). The use of the specific climate sensitivity is that it can give alinear prediction

T0 = T + Sforcing,slow (�Rforcing +�Rslow) . (3) eq:linpred

For a specific energy balance model including regime shifts we can explicitly calculate ECSfor the di↵erent regimes, see section 3. We note that several other authors have highlightedthe need to improved notes of ECS: these include [7] who propose to use a measure-basedapproach to understand climate sensitivity and [11] who consider conditional climate sensi-tivities constrained by temperature, coupled with resilience measures for switching to otherregimes.

2.1 Observation of the climate attractor

We consider the climate system as a high dimensional dynamical system that evolves alongtrajectories x(t) according to a smooth flow

x(t) = 't(x(0)) (4) e:climatesystem

4

http://dx.doi.org/10.1038/nature11574


Last 800 kyr: State dependent feedbacks

8

Friedrich et al. Sci. Adv. (2016) ↖

(with an ensemble range of 3.42 to 6.40 K) when simulated by theCMIP5 ensemble mean.

Impact of uncertaintiesThe radiative forcing and global mean SAT reconstruction are bothsubject to considerable uncertainties that affect the estimate of Sand thus the resulting global warming projection. Dust changesare the largest contributor to the uncertainty in global radiativeforcing. Reported values for the glacial-interglacial amplitude ofthe global mean dust forcing range from ~0.33 W/m2 (29) to morethan 3 W/m2 [see discussion by Köhler et al. (1)]. Here, we assumea published value of 1.9 W/m2 that is prone to large uncertainties.Regarding estimates of global mean SAT change, there appears tobe a reasonable agreement between different studies with respect tothe temporal evolution (fig. S5). However, the correct PI-LGM am-plitude of global mean SAT change remains less constrained. Pub-lished amplitudes range from relatively moderate values of 3.00 K(90% probability range of 1.7 to 3.7 K) (3) and 4.00 ± 0.80 K (21) tovalues of 5.80 ± 1.4 K (18) and 6.20 K (90% probability range of4.60 to 8.30 K) (19). To calculate the effect of uncertainties in tem-perature reconstructions on our estimate of Swarm and the resultingglobal warming projection, we scale our global mean SAT reconstruc-tion to other reported values. Furthermore, we use an Antarctic tempera-ture reconstruction (30) scaled by a high and a low polar amplification

factor (31). Using a PI-LGM amplitude of only 3 K (3) for our globalmean SAT reconstruction results in an Swarm of only 0.68 K W!1 m2 (or2.52 K per CO2 doubling) (table S3), which is approximately half thevalue found for our original temperature reconstruction. The associatedgreenhouse warming for the year 2100 amounts to 3.84 K. An amplitudeof 4.00 K, as derived from a multimodel ensemble in combination withglobal proxy data (21), results in a warming of 4.68 K by 2100 CE. Onthe basis of the estimates of S using scaled Antarctic temperature re-construction, the global warming at the end of the 21st century amountsto 6.32 and 4.87 K for the low (1.2) and the high (1.9) polar amplifi-cation, respectively.

Another source of uncertainty is introduced through differentwarming levels during previous interglacial periods. For example,our proxy-based temperature reconstruction exhibits an overallstronger interglacial warming than does our model-based one(Fig. 2A). As a result, Swarm derived by using only the proxy-basedglobal mean SAT reconstruction is slightly larger than the Swarm

calculated from the model-based reconstruction (table S3), eventhough the PI-LGM temperature amplitude is larger for the model-based SAT change. This highlights the need for a combined model-proxy approach for deriving SAT reconstructions.

DISCUSSIONConstraining the magnitude of future greenhouse warming is criticalfor risk assessment and adaptation strategies. Using our combinedproxy/modeling approach based on the 784-ka SST data and applyingit to the projected atmospheric CO2 concentrations and radiative for-cings results in SAT changes that overlap with the upper range of cur-rent CMIP5 RCP8.5 projections. The resulting paleodata-basedestimate of surface warming by 2100 CE is ~16% higher than theCMIP5 ensemble mean projection. Our results suggest that a globalsurface temperature increase of 4 K by 2100 CE (compared to the

Fig. 3. Sensitivity of global mean SAT anomalies to radiative forcing anomalies.Scatter diagram (circles) of reconstructed global mean SAT anomalies (K) (Fig. 2B)versus net radiative forcing anomalies (W/m2) (Fig. 2D) for the last 784,000 years.Anomalies are calculated with respect to PI values. Two-dimensional kernel den-sity estimate of paleo-SAT/radiative forcing data (blue shading). The thick dashedyellow curve represents nonlinear regression of paleo-SAT/radiative forcing data,along with uncertainty ranges (dashed black curves; see Materials and Methods).The thick cyan line represents linear regression for cold phases. The slope representsScold. The thick red line represents linear regression for warm phases. The slope rep-resents Swarm. Dashed horizontal lines denote warm (orange) and cold (blue) phasesusing 1 SD of the reconstructed global mean SAT anomalies as a separator. Cold(warm) phases are defined by SAT anomalies of <!5.12 K (>!1.66 K). The CMIP5transient model projections using the RCP8.5 forcing scenario are presented by pur-ple circles. Using Swarm (orange shading) and taking into account the ocean heatuptake efficiency, we can calculate the transient response to the RCP8.5 radiativeforcing. The resulting paleo-based projection with the corresponding uncertaintyranges is represented by cyan shading (see Materials and Methods).

Fig. 4. Future greenhouse warming projections. Ensemble mean simulatedglobal mean surface temperature (K) evolution using all models of the CMIP5 multi-model ensemble in their historical and RCP8.5 simulations (thick red line) andcorresponding uncertainty range (orange shading), along with estimates (blue) basedon Swarm, an estimate of ocean heat uptake efficiency and the RCP8.5 radiativeforcing time series. The corresponding uncertainty range is depicted as cyan shading(see Materials and Methods).

S C I ENCE ADVANCES | R E S EARCH ART I C L E

Friedrich et al. Sci. Adv. 2016;2 : e1501923 9 November 2016 4 of 11

on May 10, 2017

http://advances.sciencemag.org/

Dow

nloaded from

1818 P. Köhler et al.: State dependency of the equilibrium climate sensitivity

0

1

2

3

4

5

de

nsi

ty(-

)

0.0 0.5 1.0 1.5 2.0 2.5 3.0

S[CO2,LI] (K W-1

m2)

This study:ice cores (cold, 0-0.8 Myr)ice cores (warm, 0-0.8 Myr)

Hönisch11

B (cold, 0-2.1 Myr)

Hönisch11

B (warm, 0-2.1 Myr)

von der Heydt et al. (2014):ice cores (cold, 0-0.8 Myr)ice cores (warm, 0-0.8 Myr)

Martinez-Boti et al. (2015):ice cores (0-0.8 Myr)

11B, Pliocene (2.5-3.3 Myr)

1.05

1.56

1.07

1.51

0.98

1.34

1.01

0.91

Figure 9. Probability density function of different approaches to calculate specific equilibrium climate sensitivity, S[CO2,LI]. Results of thisstudy are based on pointwise analysis of the ice core (last 0.8Myr) and the Hönisch (last 2.1Myr) data for “cold” periods (1R[CO2,LI] <

�3.5Wm�2) and “warm” periods (1R[CO2,LI] > �3.5Wm�2). In von der Heydt et al. (2014), a similar split of the ice core data wasperformed. We show their results based on similar 1Tg as obtained here published in the Supporting Information in von der Heydt et al.(2014). Martínez-Botí et al. (2015) calculated S[CO2,LI] either for ice core data of the whole last 0.8Myr or based on �

11B for 0.8Myr of thePliocene between 2.5–3.3MyrBP. Vertical lines and labels give the mean of the different results.

Furthermore, in Fedorov et al. (2013) climate simulation re-sults have been discussed to understand which processes andmechanisms were responsible for the spatially very hetero-geneous changes observed during the last 5Myr, e.g. the in-crease in the polar amplification factor over time. Since theresults of Fedorov et al. (2013) were unable to explain allobservations, it was concluded that a combination of differ-ent dynamical feedbacks is underestimated in the climatemodels. We are not able to generate spatially explicit re-sults. However, from our analysis we could conclude that theequilibrium climate sensitivity represented by S[CO2,LI] wasa function of background climate state and probably changeddramatically between conditions with and without NorthernHemisphere land ice.The contribution of greenhouse gas radiative forcing and

of seasonally and latitudinally variable incoming solar radi-ation to the simulated global temperature anomalies of thelast eight interglacials have been analysed individually be-fore (Yin and Berger, 2012). It was found that the green-house gas forcing was the main driver of the simulated tem-perature change with the incoming solar radiation amplify-ing or dampening its signal for all but one interglacial (Ma-rine Isotope Stage (MIS) 7), with two interglacials (MIS 1and MIS 19) having variations close to zero. Furthermore,

they calculated the ECS (temperature rise for a doubling ofCO2) for the different interglacial background conditions andfound ECS to decrease with increasing background tempera-ture. A calculation of climate sensitivity for individual pointsin time has been performed before (PALAEOSENS-ProjectMembers, 2012) but has been rejected due to large uncertain-ties, mainly during interglacials since in the definition of S,one then needs to calculate the ratio of two small numbersin 1Tg and 1R[CO2,LI], which has typically a low signal-to-noise ratio. At first glance, this might seem contrary toour finding of a larger climate sensitivity during late Pleis-tocene interglacials when compared to late Pleistocene fullglacial conditions. However, as mentioned already in the pre-vious paragraph, the comparison of (palaeo)data-based cal-culations of S with ECS calculated from climate models isnot directly possible. Furthermore, in our approach we in-clude changes in land ice sheet (albedo forcing or 1R[LI]),while Yin and Berger (2012) kept ice sheets at present state.When investigating S[CO2,LI] over the whole range of climatestates (from full glacial conditions to a warm Pliocene with a(nearly) ice-free Northern Hemisphere resulting in a variableforcing term 1R[LI]), we therefore probe a completely dif-ferent climate regime, which is not directly comparable withresults obtained from simulations of interglacials only.

Clim. Past, 11, 1801–1823, 2015 www.clim-past.net/11/1801/2015/

Köhler, von der Heydt, et al. Clim. Past (2015) ↖

Eqilibrium climate sensitivity (ECS) is higher during interglacials than in cold periods.

von der Heydt et al. Curr. Clim. Change Rep. (2016) ↖

http://advances.sciencemag.org/cgi/doi/10.1126/sciadv.1501923

http://dx.doi.org/10.1007/s40641-016-0049-3


State dependent ECS from palaeoclimate data and models

9

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0.0

0.5

1.0

1.5

2.0

2.5

3.0

S[C

O2,L

I](K

W-1

m2)

-10 -5 0 5 10 15 20

T (K)

IPCC 2013 (90%CF range)

PALAEOSENS 2012 N (all@65Ma)

LP PETMfullGIGcold

warmG

GLE EE

800ka Pleist.800ka Plio. Early Paleogene

CMIP5

based on climate models:Shaffer 2016 (LP / PETM)Caballero 2013 (Paleogene: 2-32 CO2)Andrews 2012 (CMIP5 multi-model mean)

based on simpler models and data:Köhler 2016 (fullG / IG @2.1Ma)Anagnoustou 2016 (LE / EE)Köhler 2015 (coldG / warmG @2.1Ma)Martinez-Boti 2015 (800ka Pleist. / Plio.)vdHeydt 2014 (coldG / warmG @800ka)

Baatsen 2018 (LE)

von der Heydt et al. Curr. Clim. Change Rep. (2016) ↖

http://dx.doi.org/10.1007/s40641-016-0049-3


! Uncertainty from observations, model, unaccounted processes

Big uncertainties in quantification for radiative forcing.

Palaeoclimate: Big uncertainty in climate reconstruction.

! Climate dynamics:

feedback processes change with background climate!

Very high climate sensitivity:

• nonlinearities in the climate system - evidence for tipping?

Distributions of climate sensitivity - origin of uncertainty?

10


Palaeoclimate sensitivity S: trajectory on a ‘climate attractor’

11

R

TT (a) (b)

von der Heydt & Ashwin, Dyn. Stat. Clim. Syst. 1 dzx001 (2016) ↖

4 Peter Ashwin, Anna S. von der Heydt

clusions and some challenges for the future. Appendix A extends results of[36] and examines extremes of sensitivity associated with tipping points in amore realistic physics-based multi-box model of the glacial cycles by Gildorand Tziperman [17].

2 Sensitivities and the climate attractor

In order to understand variability, abrupt transitions and response to per-turbations we consider the climate system as a high-dimensional multiscalecomplex dynamical system whose evolving trajectories form a climate attrac-

tor. The ECS can be defined on this attractor and regimes or states may beidentified where a linear approximation of the response may be reasonable.Tipping points visible in the GMST will show up as large but occasional shiftsbetween di↵erent ‘climate regimes’ of the attractor, or indeed di↵erent attrac-tors. We visualise the attractor by projection onto climate variables relevantfor determining ECS, i.e. the GMST T and the radiative forcing R per unitarea [36]. Consider the energy balance model

cTdT

dt= Rforcing +Rslow +Rfast �ROLW, (1)

where the left hand side represents the rate of change of the global mean sur-face temperature T (with specific heat capacity cT ) and on the right hand sideRforcing is the (external) radiative forcing (including changes in CO2), Rslow

(Rfast) is the radiative perturbation due to all slow (fast) feedback processeswithin the climate system and ROLW is the outgoing longwave radiation, re-spectively. Following the formalism of [30], the specific climate sensitivity is

Sforcing,slow =�T

�Rforcing +�Rslow

⇡ dT

d(Rforcing +Rslow), (2)

which equals the Charney sensitivity S if �Rslow is the sum of all slow feed-back processes contributing to the ECS (and under the assumption of timescale separation). In practise, only some of the slow processes are accessiblefrom palaeoclimate records (e.g. only land ice), in which case the specific cli-mate sensitivity is only an approximation of the Charney sensitivity [30] (e.g.S[CO2,LI] is the specific climate sensitivity considering only land ice changes asslow feedback). This ECS gives a linear prediction for change in temperature:

T0 = T + Sforcing,slow (�Rforcing +�Rslow) . (3)

For a specific energy balance model including regime shifts we can explicitlycalculate ECS for the di↵erent regimes, see section 3. We note that severalother authors have highlighted the need to improved notions of ECS: thisincludes [8] who propose to use a measure-based approach to understand cli-mate sensitivity and [12] who consider conditional climate sensitivities con-strained by temperature, coupled with resilience measures for switching toother regimes.

4 Peter Ashwin, Anna S. von der Heydt

clusions and some challenges for the future. Appendix A extends results of[36] and examines extremes of sensitivity associated with tipping points in amore realistic physics-based multi-box model of the glacial cycles by Gildorand Tziperman [17].

2 Sensitivities and the climate attractor

In order to understand variability, abrupt transitions and response to per-turbations we consider the climate system as a high-dimensional multiscalecomplex dynamical system whose evolving trajectories form a climate attrac-

tor. The ECS can be defined on this attractor and regimes or states may beidentified where a linear approximation of the response may be reasonable.Tipping points visible in the GMST will show up as large but occasional shiftsbetween di↵erent ‘climate regimes’ of the attractor, or indeed di↵erent attrac-tors. We visualise the attractor by projection onto climate variables relevantfor determining ECS, i.e. the GMST T and the radiative forcing R per unitarea [36]. Consider the energy balance model

cTdT

dt= Rforcing +Rslow +Rfast �ROLW, (1)

where the left hand side represents the rate of change of the global mean sur-face temperature T (with specific heat capacity cT ) and on the right hand sideRforcing is the (external) radiative forcing (including changes in CO2), Rslow

(Rfast) is the radiative perturbation due to all slow (fast) feedback processeswithin the climate system and ROLW is the outgoing longwave radiation, re-spectively. Following the formalism of [30], the specific climate sensitivity is

Sforcing,slow =�T

�Rforcing +�Rslow

⇡ dT

d(Rforcing +Rslow), (2)

which equals the Charney sensitivity S if �Rslow is the sum of all slow feed-back processes contributing to the ECS (and under the assumption of timescale separation). In practise, only some of the slow processes are accessiblefrom palaeoclimate records (e.g. only land ice), in which case the specific cli-mate sensitivity is only an approximation of the Charney sensitivity [30] (e.g.S[CO2,LI] is the specific climate sensitivity considering only land ice changes asslow feedback). This ECS gives a linear prediction for change in temperature:

T0 = T + Sforcing,slow (�Rforcing +�Rslow) . (3)

For a specific energy balance model including regime shifts we can explicitlycalculate ECS for the di↵erent regimes, see section 3. We note that severalother authors have highlighted the need to improved notions of ECS: thisincludes [8] who propose to use a measure-based approach to understand cli-mate sensitivity and [12] who consider conditional climate sensitivities con-strained by temperature, coupled with resilience measures for switching toother regimes.

https://doi-org.proxy.library.uu.nl/10.1093/climsys/dzx001


The ‘climate attractor’

12

�R

T

A

B

C

D

Tthr

E

F

Ashwin & von der Heydt, J. Stat. Phys. (2019) ↖

http://link.springer.com/10.1007/s10955-019-02425-x



12

�R

T

A

B

C

D

Tthr

E

F Instantaneous/Local slope sensitivity

CO2 and water feedbacks. As in [11], we assume that the albedo varies with temperaturedue to changes in land-ice feedback processes: they assume there are threshold temperaturesT1 < T2 associated with changes of albedo ↵(T ) and define a function

⌃(T ) =(T � T1)

T2 � T1H(T � T1)H(T2 � T )) +H(T � T2) (16)

that switches from 0 for T < T1 to 1 for T > T2: H(T ) is approximately a Heaviside unitstep function and we use a smooth approximation H(T ) = (1 + tanh(T/✏H))/2 as in [11].We then set

↵(T ) = ↵1(1� ⌃(T )) + ↵2⌃(T )

to change smoothly from ↵1 of an ice surface (T < T1) to ↵2 of an ocean surface (T > T2)as in [11]).

Finally, we include a stochastic term to (15) that represents unresolved subgrid processeswith amplitude ⌘T :

cT dT = F (T,C)dt+ ⌘TdWT . (17) eq:EBM

The parameters used are listed in Table 1 except where specified as di↵erent. Note that thedeterministic equilibria of (15) are at F (T,C) = 0, i.e at

C = �(T ) := C0 exp

✏(T )�T 4 �Q0(1� ↵(T ))�G0

G1

�. (18) eq:EBMequil

This means there is a unique equilibrium for each T , but not necessarily for each C: inparticular as discussed in [11, 35] there are three branches of equilibria for a range of C. Wenote that from (18) we have at equilibrium that

�R[CO2] = ✏(T )�T 4 �Q0(1� ↵(T ))�G0 (19) eq:EBMDRTeqm

and for the parameters used we have bistability in the region

�20.77 Wm�2

< �R < 21.93 Wm�2, 102 ppm < C < 777 ppm, (20) eq:bistabrange

see the red line in Fig. 3. We can define the instantaneous sensitivity2 as

S =

d

dT�R[CO2]

��1

=⇥(✏0T + 4✏)�T 3 +Q0↵

0⇤�1, (21) eq:iSens

which corresponds to the slope of the tangent of the equilibrium (non-stochastic) model (seeFig. 1, point B). Note that for T � T0 we have ✏(T ) ⇡ 1�m and ↵

0 = 0 and so

S ⇡ (4�(1�m)T 3)�1 (22) eq:SfromW

while for T ⌧ T0 we have ✏(T ) ⇡ 1 and ↵0 = 0 and

S ⇡ (4�T 3)�1 (23) eq:SfromC

2This is referred to as local slope sensitivity in [19].

9





12

�R

T

A

B

C

D

Tthr

E



⌃(T ) =(T � T1)

T2 � T1H(T � T1)H(T2 � T )) +H(T � T2) (16)


↵(T ) = ↵1(1� ⌃(T )) + ↵2⌃(T )





C = �(T ) := C0 exp

✏(T )�T 4 �Q0(1� ↵(T ))�G0

G1





�20.77 Wm�2



S =

d

dT�R[CO2]

��1

=⇥(✏0T + 4✏)�T 3 +Q0↵

0⇤�1, (21) eq:iSens


0 = 0 and so

S ⇡ (4�(1�m)T 3)�1 (22) eq:SfromW


S ⇡ (4�T 3)�1 (23) eq:SfromC


9

Let us assume the current state at time t = 0 of the climate system is given by a measure�0 on phase space X (denoted by point A in Fig. 1). Note that this will always be a measurerather than a point because of lack of knowledge of sub-grid parametrized processes (e.g.[30]) but it will project onto the current values (T0, R0) = (T (x), R(x)) for all x in thesupport of �0. As time progresses, this state will spread to give a measure at time t that is

�t(A) = �('�t(A))

for any A ⇢ X (Fig. 1 shows a trajectory in black and others from the ensemble starting atA in grey). The incremental sensitivity for a time interval �t is then

S�t0 (x) =

T ('�t(x))� T0

R('�t(x))�R0(8) eq:Sinc

with distributionP(S�t

0 2 A) = �({x : S�t0 (x) 2 A}).

Over long time, if there is decay of correlations and mixing of trajectories on the climateattractor [31, 30] then in a weak sense �t ! M , and so we expect the distribution of long-term incremental sensitivities for �t ! 1 become time-independent for typical trajectorieswithin the attractor:

P(S10 2 A) = M({x : S

10 (x) 2 A}). (9) eq:S0inf

where

S10 (x) =

T (x)� T0

R(x)�R0.

Note that (7) means that the distribution of long-term sensitivities starting at (R0, T0) canbe written in terms of the geometry of the projected measure µ

P(S10 2 A) = µ({(T1, R1) : S

10,1 2 A}) (10) e:longtermsens

where we define the two-point sensitivity as

S10,1 =

T1 � T0

R1 �R0. (11) eq:S01

The distribution of long-term incremental sensitivity (10) for a generic choice of the initialclimate state suggests a time-independent notion of climate sensitivity that can be found bypicking pairs of points (R0,1, T0,1) independently distributed according to µ and evaluating(2). This was introduced in [32] but we propose a simpler notation here.

This means that for any A ⇢ R we can use µ to assign a probability to the sensitivitybeing in A:

P(S10,1 2 A) := µ⇥ µ

��(T0, R0), (T1, R1) : S

10,1 2 A

�. (12) eq:Sdist

with S10,1 defined as in (11). Note that in some sense this will give a maximal set of pos-

sibilities for the sensitivities in that it compares the observables T and R over all possibletime points and possible trajectories of the system. This is comparable to the conditionalclimate sensitivity of [11] except rather than dividing into regimes, they restrict to deviationsof temperature at most �T from T0. In the case that the sensitivity is fixed at S0, note thatS10,1 is a Dirac �-distribution centred at S0.

6

Incremental sensitivity (fixed Δt)





12

�R

T

A

B

C

D

Tthr

E



⌃(T ) =(T � T1)

T2 � T1H(T � T1)H(T2 � T )) +H(T � T2) (16)


↵(T ) = ↵1(1� ⌃(T )) + ↵2⌃(T )





C = �(T ) := C0 exp

✏(T )�T 4 �Q0(1� ↵(T ))�G0

G1





�20.77 Wm�2



S =

d

dT�R[CO2]

��1

=⇥(✏0T + 4✏)�T 3 +Q0↵

0⇤�1, (21) eq:iSens


0 = 0 and so

S ⇡ (4�(1�m)T 3)�1 (22) eq:SfromW


S ⇡ (4�T 3)�1 (23) eq:SfromC


9


�t(A) = �('�t(A))


S�t0 (x) =

T ('�t(x))� T0



0 2 A) = �({x : S�t0 (x) 2 A}).


P(S10 2 A) = M({x : S

10 (x) 2 A}). (9) eq:S0inf

where

S10 (x) =

T (x)� T0

R(x)�R0.


P(S10 2 A) = µ({(T1, R1) : S



S10,1 =

T1 � T0

R1 �R0. (11) eq:S01



P(S10,1 2 A) := µ⇥ µ

��(T0, R0), (T1, R1) : S

10,1 2 A

�. (12) eq:Sdist



6

Incremental sensitivity (fixed Δt)


�t(A) = �('�t(A))


S�t0 (x) =

T ('�t(x))� T0



0 2 A) = �({x : S�t0 (x) 2 A}).


P(S10 2 A) = M({x : S

10 (x) 2 A}). (9) eq:S0inf

where

S10 (x) =

T (x)� T0

R(x)�R0.


P(S10 2 A) = µ({(T1, R1) : S



S10,1 =

T1 � T0

R1 �R0. (11) eq:S01



P(S10,1 2 A) := µ⇥ µ

��(T0, R0), (T1, R1) : S

10,1 2 A

�. (12) eq:Sdist



6

Two-point sensitivity (all Δt)




Energy Balance model

13

0 10 20T [C]

0.45

0.5

0.55

(T)

a0 10 20

T [C]

0.35

0.4

0.45

0.5

0.55

(T)

b

0 500 1000C [ppm]

0

5

10

15

20

25

T [C

]

c-10 0 10

R [W/(m2/s)]

0

5

10

15

20

25

T [C

]

d

0 10 20T [C]

0.45

0.5

0.55

(T)

a0 10 20

T [C]

0.35

0.4

0.45

0.5

0.55

(T)

b

0 500 1000C [ppm]

0

5

10

15

20

25

T [C

]

c-10 0 10

R [W/(m2/s)]

0

5

10

15

20

25

T [C

]

d

CO2

0 10 20T [C]

0.45

0.5

0.55

(T)

a0 10 20

T [C]

0.35

0.4

0.45

0.5

0.55

(T)

b

0 500 1000C [ppm]

0

10

20

T [C

]

c-10 0 10

RCO2 [W/(m2/s)]

0

10

20

T [C

]d

wandering CO2 = additional noise




Two-point sensitivity

14

Conditional, incremental sensitivity for all delays up to 20kyr

-10 -5 0 5 10 RCO

2 [Wm-2]

-5

0

5

10

15

20

25

T [o C]

50

100

150

200

250

300

350

400

450

500

Tthr

-2 -1 0 1 20

0.1

0.2

dens

ity [a

u] W statesa

-2 -1 0 1 20

0.5

1

P(tip

W to

C) b

-2 -1 0 1 20

0.1

0.2

dens

ity [a

u] C statesc

-2 -1 0 1 2S [K/(Wm-2)]

0

0.5

1P(

tip C

to W

) d

-2 -1 0 1 20

0.1

0.2

dens

ity [a

u] W statesa

-2 -1 0 1 20

0.5

1

P(tip

W to

C) b

-2 -1 0 1 20

0.1

0.2

dens

ity [a

u] C statesc

-2 -1 0 1 2S [K/(Wm-2)]

0

0.5

1

P(tip

C to

W) d




Skewed PDFs of two-point sensitivity

15

Tipping and state-dependence of feedbacks

Only state-dependence of feedbacks

No state-dependence of feedbacks

non-constant feedback factors

non-constant feedback factors

constant feedback factors ~ Gaussian PDF

-10 -5 0 5 10

0

10

20

T [o C]

a

0 0.5 1 1.5 2 2.50

0.1

0.2

0.3

dens

ity [a

u]

d

-10 -5 0 5 10

0

10

20

T [o C]

b

0 0.5 1 1.5 2 2.50

0.01

0.02

0.03

0.04

dens

ity [a

u]e

-10 -5 0 5 10 RCO2

[Wm-2]

0

10

20

T [o C]

c

0 0.5 1 1.5 2 2.5S [K/(Wm-2)]

0

0.05

0.1

0.15

0.2

dens

ity [a

u]

f

Low albedo contrast

Standard albedo contrast

NO albedo contrast




Simple Earth System model:

Atmosphere,

Ocean,

Sea ice,

Land ice,

Ocean biogeochemistry & dynamic atmospheric pCO2,

Milankovitch forcing: insolation per box & NH land ice ablation.

Conceptual climate model

16

Gildor & Tziperman 2000, 2001, 2002




Model glacial-interglacial cycles

17




State dependent CS glacial cycles (conceptual climate model)

18

0 1 2 3 4 5 R[CO2] [Wm-2]

10

12

14

T [o C]

a20

40

60

-22 -21 -20 -19 -18 R[LI] [Wm-2]

10

12

14

T [o C]

b 2

4

6

8

-20 -18 -16 -14 R[CO2,LI] [Wm-2]

10

12

14

T [o C]

c20

40

60

Earth System Sensitivity

‘forcing’ = only CO2

Equilibrium Climate Sensitivity

‘forcing’ = CO2 + Land Ice

ESS corrected for slow feedbacks

0 1 2 3 4 5 R[CO2] [Wm-2]

10

12

14

T [o C]

a20

40

60

-22 -21 -20 -19 -18 R[LI] [Wm-2]

10

12

14

T [o C]

b 2

4

6

8

-20 -18 -16 -14 R[CO2,LI] [Wm-2]

10

12

14

T [o C]

c20

40

60



Dr. Anna von der Heydt 19

-1 0 1 2 30

0.2

0.4

dens

ity [a

u]

W statesa

-1 0 1 2 30

0.5

1

P(tip

W to

C) b

-1 0 1 2 30

0.01

0.02

0.03

dens

ity [a

u]

C statesc

-1 0 1 2 3S[CO2,LI] [K/(Wm-2)]

0

0.5

1

P(tip

C to

W) d

0 1 2 3 4 5 R[CO2] [Wm-2]

10

12

14

T [o C]

a20

40

60

-22 -21 -20 -19 -18 R[LI] [Wm-2]

10

12

14

T [o C]

b 2

4

6

8

-20 -18 -16 -14 R[CO2,LI] [Wm-2]

10

12

14

T [o C]

c20

40

60

As in energy balance model:

High values of S[CO2,LI] correspond to high probability of regime (cold/warm) transition


Two-point sensitivity & probability of tipping (conceptual climate model)



! Climate sensitivity depends on the background climate state:

non-constant fast feedback processes,

multiple equilibrium or oscillatory states (‘tipping’).

! ‘Flavours’ of (palaeo)climate sensitivity on the ‘climate attractor:

instantaneous S: available from underlying model (‘nearest equilibrium’),

incremental S: fixed delay Δt,

two-point S: all delays, two points on attractor.

! Nonlinearities lead to skewed PDFs of measured climate sensitivity

Extremes of climate sensitivity seem to relate to high probability of tipping.

Conclusions

20

uncertainty quantification of climate sensitivity: state ...€¦ · uncertainty quantification of...

Documents